AI and artists

Where bitwerk stands on AI: AI may help write the code and may help find, sort and explain music. It may never make music, voices, images or any other art, and bitwerk may not be used to make it. This page gives our reasons, says what bitwerk is for, and sets the binding rules, the Artist Protection Act. Owner decision of 2026-10-20.

Our stance in short

	Where we stand
AI for code	Yes. bitwerk is written with the help of AI coding assistants. A person decides what is built, and every change is reviewed and tested before it lands. The code will be published: free for personal use, free for research and teaching on request, and open source under the AGPL four years after each release.
AI for information	Yes, with rules. Searching, filtering and improving information about music that exists: facts with sources, free photos, tags, similar tracks, transitions between real recordings. Off until an admin turns it on.
AI art	No. No AI-generated music, vocals, voices, lyrics, images, covers, videos or any other art in bitwerk, and no use of bitwerk to make them.

Why we say no to AI art

It is built on artists’ work without their consent. Generative music and image models learn from recordings, songs, lyrics and pictures that people made, mostly without asking, naming or paying them. In November 2025 the Munich Regional Court found that a language model which stores and reproduces song lyrics infringes the songwriters’ rights (GEMA v. OpenAI, LG München I, 42 O 14139/24, not final).
It competes with the people it learned from. In July 2026 Deezer reported that fully AI-generated tracks made up more than half of its daily deliveries on peak days, nearly 90,000 a day, while drawing only 1–3 % of the streams, and that up to 85 % of the streams on them in 2025 were fraudulent (Deezer, 2026-07-21). That flood takes attention and royalties away from musicians.
A voice and a face belong to a person. Voice clones and deepfakes take what is most personal to a singer, a DJ or a presenter, and can make them say what they never said.
Listeners deserve to know who made what they hear. bitwerk promises that what reaches the DAC is the file, unaltered, and that the screen says so. Generated content passed off as music by people breaks the same trust from the other side.
The people who use bitwerk live from human craft. Studios, mastering engineers, DJs, clubs, bars, listening rooms and events are what bitwerk is built for. A tool for them should not help replace them.
The law is not settled. The training and the output of generative models are in court in several countries. A business that runs bitwerk should not carry that risk because of a feature in its music system.

Why AI for code is fine with us

Code is a tool. It is judged by whether it works, and every change is reviewed and tested by a person who answers for it, as with code written by hand. It does not imitate an artist’s voice, face or style, and it does not compete with musicians for listeners.
We give the result back. The code will be published for everyone to read and use: free for individuals, free for research and teaching on request, and open source under the AGPL after four years (licence).
We know the debate. Code models learn from public code, too. We use them as tools under human review, and we draw the line where a model would make or imitate a work of art.

What bitwerk is for

bitwerk brings recorded music from the file to the speaker unaltered, in rooms where sound is the job: listening rooms and dealers’ demo rooms, studios, bars, clubs, events, and at home. It runs on your own hardware without a cloud account, shows what reaches the DAC, keeps rooms in sync, and gives studios and venues their tools in one system. It is made for the people who make, mix, play and listen to music.

AI serves that goal and nothing else: it helps find, sort and understand music that people made, and never replaces them. How each AI feature works, what it sends where and when it comes is in the AI plan.

The Artist Protection Act

bitwerk’s own rules to protect artists from AI. They bind the project: what it builds, what it accepts from contributors and how its features work. For users they are bitwerk’s terms of use, and §3 is part of bitwerk’s licence (§5). They are not a law and not legal advice.

1. What AI in bitwerk may do

AI in bitwerk has three jobs: search, filter and improve information about recordings that exist.

Job	Examples	Milestones
Search	describe how a song sounded and find it, “Sounds like this”, find free photos of an artist	M26, M88, M87
Filter	autoplay and smart shuffle from your own library, AI tags as filters, demo lists for dealers	M89, M88, M92
Improve information	song and artist facts with sources, cleaning up messy tags, translating a sourced bio, explaining a device problem	M87, M92

Two features come close to the music and stay on the same side of the line:

DJ transitions (M90) play one existing recording into the next, with volume, EQ and tempo, as a DJ does by hand. Nothing is generated, recorded or saved.
The DJ voice (M91) reads facts with sources in a licensed synthetic voice that imitates no one. It is labelled as AI and off by default.

2. What bitwerk never does

No generated music: no songs, loops, samples, continuations, remixes, cover versions, “in the style of” tracks or style transfer made by a model, and no model in bitwerk that could make them.
No voice of a real person: no voice clones, no singing synthesis, no imitation of artists, DJs or presenters. Speech uses only catalogued voices with a licence for commercial use that copy no real person; there is no upload of reference audio.
No generated images or videos: no AI photos, portraits, covers or videos, and no AI editing of artist photos or covers beyond cropping and scaling. Artist photos are free photos with their author and licence, or your own uploads.
No generated lyrics: a model neither writes, completes, translates nor explains lyrics.
Nothing put in an artist’s mouth: AI text states facts with sources, is labelled as AI and never invents quotes, statements or intentions of an artist.
No training: bitwerk never trains or fine-tunes a model on music, voices, lyrics, covers, photos or listening histories. Its built-in models are pretrained and only recognise and compare (sound fingerprints, tags).
Audio never goes to an AI provider: sound is analysed only on your own server.
Separation only live: if stem separation comes after v2.0 (vocals off for karaoke), it works only while the music plays, never writes stems to a file and never sends them anywhere.

3. What nobody may use bitwerk for

Nobody may use bitwerk, its software, interfaces, streams, analysis or AI features to:

collect training data for a model that generates music, voices, lyrics, images or videos: no bulk downloading or recording of the library or of streams for that purpose, and no export of analysis, sound fingerprints or AI tags for it;
train, fine-tune or test such a model;
make voice clones, deepfakes or other imitations of artists, or AI art in an artist’s style;
present generated content as an artist’s work, or remove credits, sources, licences or AI labels.

4. How bitwerk keeps to it

Fixed tasks, no open chat: the AI service sends only the prompts and answer formats that bitwerk defines for each feature. No interface passes a free prompt to a model or returns generated audio, images or video.
No capability for AI art in the code: the AI providers in bitwerk know text, recognising images, sound fingerprints and the DJ voice. Generating music or images and cloning voices do not exist as capabilities, and a change that adds one is not accepted.
Fingerprints stay inside: no interface hands out the raw sound fingerprints (embeddings); a search answers with tracks and scores. The backup holds them only to restore the same server.
Streams only for players and people: audio goes only to paired players and signed-in users, with a token or a signed, expiring address (architecture).
Visible use: every request to an AI provider is in the audit log, and the AI interfaces are rate-limited per account.
Credits stay: facts show their sources and photos their author and licence; composer, lyricist and performer credits come from the files’ tags and MusicBrainz, never from a model.
Per-artist switch: an admin can turn AI off for an artist: no AI text, no AI photo search, no DJ voice line about them.
Labels: AI text carries an AI label in the interface and a marker in the API; the DJ voice says that it is synthetic (EU AI Act, Article 50).
Spotify: nothing from Spotify goes into a model, as Spotify’s Developer Policy requires (AI plan §4.10).

5. Licence and legal status

In the licence (owner decision 2026-10-20): the Additional Use Grant of the Business Source License 1.1 in LICENSE excludes the uses of §3, and no commercial licence covers them either:

Production use for any of the following purposes is not covered by this grant, and the Licensor does not offer a commercial license for it: (a) collecting training data for, training, fine-tuning or testing a machine-learning model that generates music, voices, lyrics, images or video; (b) creating voice clones, deepfakes or other imitations of performers; (c) presenting content generated by such a model as the work of an artist.
The BSL lets the licensor narrow that grant, but not restrict the base rights to copy, change and test the code (non-production use); those stay as they are.
Four years after each release the code becomes open source under the AGPL, which allows no restrictions of use. From then on §2 and §4 protect artists through what bitwerk builds, not through the licence.
The protocol (proto/, MIT or Apache 2.0) and the documentation (CC BY 4.0) stay open.
A lawyer checks this wording before the first release with AI features.

6. Changes

Changes to this act need the owner’s decision and are dated on this page.